Week of August 30: Check Its Work

Check Its Work

Last week you could point Supanova at a repository and get back the company. This week is about what happens once that company starts working.

The theme this week is that the platform stopped guessing.

What's New

It runs until the work exists

A task no longer ends because its context or token strategy ran out. Any atom working on a task will keep working, saving its place, and picking itself back up until there is real evidence that what you asked for was delivered and checked.

Open the thing you were handed

Standalone tasks now have an External Results section. Every Google Sheet or other artifact the work produced sits there behind an Open deliverable link. One click, new tab.

Older work was normalized too, so deliverables from before this week open the same way.

Your comment survives the restart

A comment saved, an atom that worked out its reply, and a server that rolled over before the reply was written. It happened once. Once is enough.

Replies are now queued in the database, claimed, and retried until they land. A restart no longer costs you the answer, and recovery can't double-post, because a reply is tied to the comment that asked for it. Persona, memory, revisions, threading, and live updates behave exactly as before.

New efficiency metrics

Atom cards used to show a simple efficiency figure.

Efficiency is now measured from what an atom actually delivered over the last ninety days, against the other atoms in your workspace doing comparable work.

Whatever model you picked, the platform behaves

Pick a model in the Lab and it now runs through the same tool machinery every platform model uses.

Seven guides, in the Help Center

New walkthroughs for choosing a model, Adaptive Compute, dynamic model selection, the CFO Atom, and the difference between API keys and Catalyst Bonds.

The Pods and Workflows guides were rebuilt around the questions people actually arrive with: will it repeat reliably, can I simulate it first, is it ready, how do I verify it, what keeps it safe.

Under the Hood

Adaptive Compute was rebuilt. It used to act like a governor, capping what work could use. It is now an estimator: it watches real demand without capping it, and it learns from everything a task consumed rather than a slice of it.

Separately, the platform had been finishing provider calls and recording nothing about how they went. Your ledger was correct and authoritative throughout; the learning signal was the blank part.

Outcome and settlement are now written in a single transaction, so the record lands and the system learns from its own work.

Sixty-five weeks in. You built it, you talk to it, you trust it, and now you can check its work. See you next Sunday.